Estimasi Fungsi Intensitas Spatial Point Process Dengan Light Gradient Boosting Machine (LightGBM) Dan Extreme Gradient Boosting (XGBoost): Studi Kasus Analisis Gempa Bumi Pulau Sumatera

Tasbihi, Naufal Luthfan (2025) Estimasi Fungsi Intensitas Spatial Point Process Dengan Light Gradient Boosting Machine (LightGBM) Dan Extreme Gradient Boosting (XGBoost): Studi Kasus Analisis Gempa Bumi Pulau Sumatera. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6003242014-Master_Thesis.pdf Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Estimasi fungsi intensitas spatial point process dengan pendekatan nonparametrik membutuhkan lebih sedikit asumsi dan fleksibel karena tidak dibatasi oleh suatu fungsi yang didefinisikan. Namun, metode nonparametrik seperti KIE dan GP memiliki kendala yaitu terbatas dalam jumlah kovariat dan membutuhkan waktu komputasi yang lama. Pendekatan nonparametrik berbasis gradient boosting seperti XGBoostPP menunjukkan performa yang sangat baik dan mampu mengakomodasi jumlah kovariat yang banyak. Akan tetapi XGBoostPP memiliki kekurangan yaitu waktu komputasi yang lama karena metode pencarian split node nya yang kurang efisien. Oleh karena itu, pada penelitian ini akan diusulkan metode berbasis gradient boosting yang lain yaitu LightGBM. LightGBM memiliki karakteristik yang hampir sama dengan XGBoost namun dengan algoritma yang lebih efisien dalam pencarian split node dan konvergensi disebabkan oleh algoritma Gradient-Based One Side Sampling (GOSS) dan Exclusive Feature Bundling (EFB), menawarkan pendekatan yang lebih efisien dalam hal komputasi. Penelitian ini bertujuan untuk mengaplikasikan dan membandingkan LightGBM, XGBoostPP, dan pendekatan nonparametrik lain dengan studi kasus analisis gempa bumi di Pulau Sumatera. Evaluasi yang digunakan adalah akurasi prediksi menggunakan log-likelihood, prediksi jumlah kejadian, dan efisiensi komputasi. Hasil penelitian menunjukkan LightGBMPP unggul signifikan dalam kecepatan dengan waktu komputasi 118,903 detik, sekitar dua kali lebih cepat dibandingkan XGBoostPP (252,889 detik). Namun, terdapat trade-off akurasi di mana XGBoostPP memiliki goodness-of-fit yang lebih baik, ditunjukkan oleh nilai Poisson Log-Likelihood 7278,386 dan MSE Spasial 0,038, dibandingkan LightGBMPP (Log-Likelihood 7094,195; MSE Spasial 0,068). Kedua model mengidentifikasi jarak ke zona subduksi dan sesar sebagai faktor paling berpengaruh, dengan LightGBMPP yang lebih efisien meski lebih sensitif terhadap pengaturan hyperparameter.
===================================================================================================================================
The estimation of the intensity function of a spatial point process through a nonparametric approach necessitates a reduced set of assumptions and offers greater flexibility due to its independence from a defined function. Nonetheless, nonparametric methods, including KIE and GP, are constrained by the number of available covariates and necessitate extended computation times. Gradient boosting-based nonparametric approaches, such as XGBoostPP, demonstrate exceptional performance and possess the capacity to accommodate a substantial number of covariates. However, XGBoostPP is subject to a disadvantage in that it necessitates a protracted computation time due to its inefficient split node search method. Consequently, this study proposes an additional gradient boosting-based method, namely LightGBM. LightGBM exhibits characteristics analogous to those of XGBoost, yet it employs a more efficient algorithm in the context of split node search and convergence. This efficiency is attributed to the implementation of the Gradient-Based One Side Sampling (GOSS) and Exclusive Feature Bundling (EFB) algorithms, which collectively provide a more efficient approach in terms of computation. The objective of this study is to implement and evaluate LightGBM, XGBoostPP, and other nonparametric methodologies through a case study of earthquake analysis on the island of Sumatra. The evaluation employed included prediction accuracy as measured by log-likelihood, the prediction of the number of events, and computational efficiency. The experimental results demonstrate that LightGBMPP exhibits a marked superiority in terms of processing speed, with a computation time of 118.903 seconds, which is approximately twice as fast as the XGBoostPP algorithm (252.889 seconds). However, this approach is associated with an accuracy trade-off, wherein XGBoostPP demonstrates superior goodness-of-fit, as evidenced by a Poisson Log-Likelihood value of 7278.386 and a Spatial MSE of 0.038, in comparison to LightGBMPP (Log-Likelihood 7094.195; Spatial MSE 0.068). The two models identify distance to subduction zones and faults as the most influential factors. LightGBMPP is more efficient despite being more sensitive to hyperparameter settings.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Exclusive Feature Bundling, Gradient-Based One Side Sampling, Spatial Point Process,
Subjects:	H Social Sciences > HA Statistics > HA30.6 Spatial analysis
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49101-(S2) Master Thesis
Depositing User:	Naufal Luthfan Tasbihi
Date Deposited:	22 Jan 2026 09:32
Last Modified:	22 Jan 2026 09:32
URI:	http://repository.its.ac.id/id/eprint/130115

Actions (login required)

View Item